Successful integration of deep neural networks (DNNs) or deep learning (DL) has resulted in breakthroughs in many areas. However, deploying these highly accurate models for data-driven, learned, automatic, and practical machine learning (ML) solutions to end-user applications remains challenging. DL algorithms are often computationally expensive, power-hungry, and require large memory to process complex and iterative operations of millions of parameters. Hence, training and inference of DL models are typically performed on high-performance computing (HPC) clusters in the cloud. Data transmission to the cloud results in high latency, round-trip delay, security and privacy concerns, and the inability of real-time decisions. Thus, processing on edge devices can significantly reduce cloud transmission cost. Edge devices are end devices closest to the user, such as mobile phones, cyber–physical systems (CPSs), wearables, the Internet of Things (IoT), embedded and autonomous systems, and intelligent sensors. These devices have limited memory, computing resources, and power-handling capability. Therefore, optimization techniques at both the hardware and software levels have been developed to handle the DL deployment efficiently on the edge. Understanding the existing research, challenges, and opportunities is fundamental to leveraging the next generation of edge devices with artificial intelligence (AI) capability. Mainly, four research directions have been pursued for efficient DL inference on edge devices: 1) novel DL architecture and algorithm design; 2) optimization of existing DL methods; 3) development of algorithm–hardware codesign; and 4) efficient accelerator design for DL deployment. This article focuses on surveying each of the four research directions, providing a comprehensive review of the state-of-the-art tools and techniques for efficient edge inference.

Modern healthcare is transforming from hospital-centric to individual-centric systems. Emerging implantable and wearable medical (IWM) devices are integral parts of enabling affordable and accessible healthcare. Early disease diagnosis and preventive measures are possible by continuously monitoring clinically significant physiological parameters. However, most IWM devices are battery-operated, requiring replacement, which interrupts the proper functioning of these devices. For the continuous operation of medical devices for an extended period of time, supplying uninterrupted energy is crucial. A sustainable and health-compatible energy supply will ensure the high-performance real-time functioning of IWM devices and prolong their lifetime. Therefore, harvesting energy from the human body and ambient environment is necessary for enduring precision healthcare and maximizing user comfort. Energy harvesters convert energy from various sources into an equivalent electrical form. This paper presents a state-of-the-art comprehensive review of energy harvesting techniques focusing on medical applications. Various energy harvesting approaches, working principles, and the current state are discussed. In addition, the advantages and limitations of different methods are analyzed and existing challenges and prospects for improvement are outlined. This paper will help with understanding the energy harvesting technologies for the development of high-efficiency, reliable, robust, and battery-free portable medical devices.

The adverse glycemic events triggered by the inaccurate insulin infusion in Type I diabetes (T1D) can lead to fatal complications. Predicting blood glucose concentration (BGC) based on clinical health records is critical for control algorithms in the artificial pancreas (AP) and aiding in medical decision support. This paper presents a novel deep learning (DL) model incorporating multitask learning (MTL) for personalized blood glucose prediction. The network architecture consists of shared and clustered hidden layers. Two layers of stacked long short-term memory (LSTM) form the shared hidden layers that learn generalized features from all subjects. The clustered hidden layers comprise two dense layers adapting to the gender-specific variability in the data. Finally, the subject-specific dense layers offer additional fine-tuning to personalized glucose dynamics resulting in an accurate BGC prediction at the output. OhioT1DM clinical dataset is used for the training and performance evaluation of the proposed model. A detailed analytical and clinical assessment have been performed using root mean square (RMSE), mean absolute error (MAE), and Clarke error grid analysis (EGA), respectively, which demonstrates the robustness and reliability of the proposed method. Consistently leading performance has been achieved for 30- (RMSE = 16.06 $\pm$ 2.74, MAE = 10.64 $\pm$ 1.35), 60- (RMSE = 30.89 $\pm$ 4.31, MAE = 22.07 $\pm$ 2.96), 90- (RMSE = 40.51 $\pm$ 5.16, MAE = 30.16 $\pm$ 4.10), and 120-minute (RMSE = 47.39 $\pm$ 5.62, MAE = 36.36 $\pm$ 4.54) prediction horizon (PH). In addition, the EGA analysis confirms the clinical feasibility by maintaining more than 94 % BGC predictions in the clinically safe zone for up to 120-minute PH. Moreover, the improvement is established by benchmarking against the state-of-the-art statistical, machine learning (ML), and deep learning (DL) methods.

Lung cancer is the abnormal growth of cells in the lung causing severe risk to human health because lung has a connected network of blood vessels and lymphatic channels vulnerable to metastasis. The survival rate of lung cancer greatly depends on the early detection and staging of the lung tumor. Computed Tomography (CT) images are widely used for lung tumor detection as they provide information about different parts of the lung. The accuracy of detecting tumor shape, location and volume plays a vital role in successful diagnosis and treatment of tumors. So, tumor detection and representation from three dimensional (3D) CT images are necessary to measure the tumor size properly and to detect the exact location of the tumor. Researchers introduced different methods for tumor nodule detection from two dimensional (2D) CT images, hence, the problems associated with such methods is the separation of lung from background and the differentiation of tumor from airways. This paper presents the use of neighborhood and connectivity properties of the CT image pixels to overcome these problems. In this work, several morphological processing used to remove background, noises and airways from the lung CT image and then the k-means clustering algorithms based segmentation used for lung tumor detection. The volumetric analysis of the lung nodules for the prediction of tumor stages performed according to the Tumor Nodule Metastasis (TNM) classification proposed by the World Health Organization (WHO). TNM for lung cancer uses the dimension, spreading and metastasis properties of the tumor. The lung tumor segmentation performed on a dataset named SPIE-AAPM Lung CT Challenge dataset collected from Washington University in St. Louis, which includes 22,489 CT images of 70 patients. The proposed scheme for 3D CT image segmentation and classification provides better performance in terms of accurate tumor detection and visualization of variant size, shape and location, less computation time and proper classification of tumor stages with an accuracy of 95.68%.

This work comprehensively investigates the systematic integration of highly relevant life events and physiological parameters with continuous glucose monitoring (CGM) data to understand the synergy in glucoregulatory systems. Multitask learning (MTL) is employed in learning from numerous subjects while tailoring features in patient-specific layers. Systematic combinations of inputs are fed to the system providing personalized blood glucose level (BGL) predictions at multi-step prediction horizons (PHs) as output. Three cutting-edge long-short-term memory (LSTM) networks are adapted in the shared layers of MTL architecture. Moreover, PHs are varied with 30-minute intervals up to 120 minutes to identify the long-term effects of relevant input features on BGL prediction suggesting optimal deep learning (DL) architecture. The empirical result demonstrates that the most relevant features for BGL prediction are glucose, bolus insulin, and carbohydrate estimate from meals, while exercise and basal insulin rate have momentary effects. The best predictive root means square error (RMSE) achieved are 16.06±2.74, 30.89±4.31, 40.51±5.16, and 47.31±5.78 mg/dL for 30-, 60-, 90-, and 120-min PH, respectively, maintaining 94.06±3.08 % predictions in clinically safe zones (A+B) in Clarke error grid analysis (EGA). The insights learned from the experiments will assist in selecting appropriate DL models, features, and timelines based on specific needs, with significant promise in improving T1D management through better therapeutic and lifestyle modification.